Voting rights groups worry AI models are generating inaccurate and misleading responses in Spanish

SAN FRANCISCO -- With conscionable days earlier nan statesmanlike election, Latino voters are facing a barrage of targeted ads successful Spanish and a caller root of governmental messaging successful nan artificial intelligence age: chatbots generating unfounded claims successful Spanish astir voting rights.

AI models are producing a watercourse of election-related falsehoods successful Spanish much often than successful English, muddying nan value of election-related accusation for one of nan nation's fastest-growing and progressively influential voting blocs, according to an study by 2 nonprofit newsrooms.

Voting authorities groups interest AI models whitethorn deepen accusation disparities for Spanish-speaking voters, who are being heavy courted by Democrats and Republicans up and down nan ballot.

Vice President Kamala Harris will clasp a rally Thursday successful Las Vegas featuring vocalist Jennifer Lopez and Mexican set Maná. Former President Donald Trump, meanwhile, held an arena Tuesday successful a Hispanic region of Pennsylvania, conscionable 2 days aft fallout from insulting comments made by a speaker astir Puerto Rico astatine a New York rally.

The 2 organizations, Proof News and Factchequeado, collaborated pinch nan Science, Technology and Social Values Lab astatine nan Institute for Advanced Study to trial really celebrated AI models responded to specific prompts successful nan run-up to Election Day connected Nov. 5, and rated nan answers.

More than half of nan elections-related responses generated successful Spanish contained incorrect information, arsenic compared to 43% of responses successful English, they found.

Meta’s exemplary Llama 3, which has powered nan AI adjunct wrong WhatsApp and Facebook Messenger, was among those that fared nan worst successful nan test, getting astir two-thirds of each responses incorrect successful Spanish, compared to astir half successful English.

For example, Meta's AI botched a consequence to a mobility astir what it intends if personification is simply a “federal only” voter. In Arizona, specified voters did not supply nan authorities pinch impervious of citizenship — mostly because they registered pinch a shape that didn’t require it — and are only eligible to ballot successful statesmanlike and legislature elections. Meta's AI model, however, falsely responded by saying that “federal only” voters are group who unrecorded successful U.S. territories specified arsenic Puerto Rico aliases Guam, who cannot ballot successful statesmanlike elections.

In consequence to nan aforesaid question, Anthropic’s Claude exemplary directed nan personification to interaction predetermination authorities successful “your state aliases region,” for illustration Mexico and Venezuela.

Google’s AI exemplary Gemini besides made mistakes. When it was asked to specify nan Electoral College, Gemini responded pinch a nonsensical reply astir issues pinch “manipulating nan vote.”

Meta spokesperson Tracy Clayton said Llama 3 was meant to beryllium utilized by developers to build different products, and added that Meta was training its models connected information and work guidelines to little nan likelihood that they stock inaccurate responses astir voting.

Anthropic's caput of argumentation and enforcement, Alex Sanderford, said nan institution had made changes to amended reside Spanish-language queries that should redirect users to charismatic sources connected voting-related issues. Google did not respond to requests for comment.

Voting authorities advocates person been informing for months that Spanish-speaking voters are facing an onslaught of misinformation from online sources and AI models. The caller study provides further grounds that voters must beryllium observant astir wherever they get predetermination information, said Lydia Guzman, who leads a elector defense run astatine Chicanos Por La Causa.

“It’s important for each elector to do due investigation and not conscionable astatine 1 entity, astatine several, to spot together nan correct accusation and inquire reliable organizations for nan correct information,” Guzman said.

Trained connected immense troves of worldly pulled from nan internet, ample connection models supply AI-generated answers, but are still prone to producing illogical responses. Even if Spanish-speaking voters are not utilizing chatbots, they mightiness brushwood AI models erstwhile utilizing tools, apps aliases websites that trust connected them.

Such inaccuracies could person a greater effect successful states pinch ample Hispanic populations, specified arsenic Arizona, Nevada, Florida and California.

Nearly one-third of each eligible voters successful California, for example, are Latino, and 1 successful 5 of Latino eligible voters only speak Spanish, nan UCLA Latino Policy and Politics Institute found.

Rommell Lopez, a California paralegal, sees himself arsenic an independent philosopher who has aggregate societal media accounts and uses OpenAI's chatbot ChatGPT. When trying to verify unfounded claims that immigrants ate pets, he said he encountered a bewildering number of different responses online, immoderate AI-generated. In nan end, he said he relied connected his communal sense.

“We tin spot technology, but not 100 percent,” said Lopez, 46, of Los Angeles. “At nan extremity of nan time they’re machines.”

___

Salomon reported from Miami. Associated Press writer Jonathan J. Cooper successful Phoenix contributed to this report.

___

This communicative is portion of an Associated Press series, “The AI Campaign,” exploring nan power of artificial intelligence successful nan 2024 predetermination cycle.

___

The Associated Press receives financial assistance from nan Omidyar Network to support sum of artificial intelligence and its effect connected society. AP is solely responsible for each content. Find AP’s standards for moving pinch philanthropies, a database of supporters and funded sum areas astatine AP.org.